Dogs vs. Cats¶
一、数据预处理¶
1、定义函数¶
In [1]:
# -*- coding=UTF-8 -*-
import os
import shutil
from tqdm import tqdm
from urllib.request import urlretrieve
IF_MKDIR = 0 # 创建的文件夹是否覆盖原来的
def rmdir_mkdir(dirname):
'''创建文件夹，若存在删除，重新创建
Args:
dirname:文件夹名
Returns:
None
'''
if os.path.exists(dirname) and IF_MKDIR:
shutil.rmtree(dirname) # 若存在则删除该文件夹（包括里面的内容
if IF_MKDIR:
os.mkdir(dirname)
class DLProgress(tqdm):
'''进度条类，继承自tqdm
'''
last_block = 0
def hook(self, block_num=1, block_size=1, total_size=None):
self.total = total_size
self.update((block_num - self.last_block) * block_size)
self.last_block = block_num
2、下载数据、解压数据包¶
In [2]:
import zipfile
if not os.path.isfile('train.zip'):
with DLProgress(unit='B', unit_scale=True, miniters=1, desc='Andimal Dataset') as pbar:
urlretrieve('https://www.kaggle.com/c/dogs-vs-cats/download/train.zip',
'train.zip',
pbar.hook)
if not os.path.isfile('train.zip'):
with DLProgress(unit='B', unit_scale=True, miniters=1, desc='Andimal Dataset') as pbar:
urlretrieve('https://www.kaggle.com/c/dogs-vs-cats/download/test1.zip',
'test1.zip',
pbar.hook)
if not os.path.isdir("train"):
with zipfile.ZipFile('train.zip') as zipp:
zipp.extractall()
zipp.close()
if not os.path.isdir("test1"):
with zipfile.ZipFile('test1.zip') as zipp:
zipp.extractall()
zipp.close()
3、导入Keras包¶
In [3]:
from keras.models import *
from keras.layers import *
from keras.applications import *
from keras.preprocessing.image import *
from keras.callbacks import *
from keras.callbacks import ReduceLROnPlateau, Callback, CSVLogger, History, ModelCheckpoint, EarlyStopping
4、图像载入、尺寸变换¶
In [4]:
import numpy as np
from tqdm import tqdm
import cv2
np.random.seed(42)
class Pictures(object):
def __init__(self,n=25000):
"""构造函数
# Arguments
n: 图像数量
# Return
None
"""
self.pic_num = n
self.shape = None
self.X = None
self.y = None
def resize(self, shape=(224,224)):
"""图像归一化
img = Pictures()
img.resize()
# Arguments
shape: 规划化后图像尺寸
# Return
None
"""
self.shape = shape
self.X = np.zeros((self.pic_num, self.shape[0], self.shape[1], 3), dtype=np.uint8)
self.y = np.zeros((self.pic_num, 1), dtype=np.uint8)
for i in tqdm(range(int(self.pic_num/2))):
self.X[i] = cv2.resize(cv2.imread('train/cat.%d.jpg' %i), (self.shape[0], self.shape[1]))
self.X[i+int(self.pic_num/2)] = cv2.resize(cv2.imread('train/dog.%d.jpg' %i), (self.shape[0], self.shape[1]))
self.y[int(self.pic_num/2):] = 1
In [7]:
# 绘制图形
import matplotlib.pyplot as plt
aninal_num = [12500,12500]
plt.bar([0,1],
aninal_num,
color=['r','b'],
tick_label =['Cats','Dogs'],
label = ['12500','12500'])
# plt.xticks([0,1],('Cats','Dogs'))
plt.text(0,12500,"12500")
plt.text(1,12500,"12500")
plt.show()
In [8]:
import matplotlib.pyplot as plt
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
img_224 = Pictures()
img_224.resize(shape=(224,224))
img_299 = Pictures()
img_299.resize(shape=(299,299))
In [9]:
# OpenCV版本
idx = np.random.randint(0,img_299.X.shape[0],size = 10)
fig, axes = plt.subplots(2,5,sharex=True ,sharey=True ,figsize=(20,8))
for i, ax in zip(idx,axes.flatten()):
ax.imshow(img_299.X[i][:,:,::-1],aspect='equal')
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
plt.subplots_adjust(wspace=0,hspace=0)
5、将数据分为训练集、验证集¶
In [10]:
from sklearn.model_selection import train_test_split
X_train_224, X_valid_224, y_train_224, y_valid_224 = train_test_split(img_224.X, img_224.y, test_size=0.2)
In [11]:
# 绘制图形
import matplotlib.pyplot as plt
%matplotlib inline
aninal_num = [len(X_train_224),len(y_train_224),len(X_valid_224),len(y_valid_224)]
plt.bar([0,1,2,3],
aninal_num,
align='center',
color=['r','r','b','b'],
tick_label =['X_train','y_train','X_valid','y_valid'],
label = [str(len(X_train_224)),str(len(y_train_224)),str(len(X_valid_224)),str(len(y_valid_224))])
plt.text(0,len(X_train_224)*1.02,str(len(X_train_224)))
plt.text(1,len(y_train_224)*1.02,str(len(y_train_224)))
plt.text(2,len(X_valid_224)*1.02,str(len(X_valid_224)))
plt.text(3,len(y_valid_224)*1.02,str(len(y_valid_224)))
plt.show()
二、使用Keras预训练模型测试¶
1、VGG19¶
In [11]:
from keras.applications.vgg19 import VGG19, preprocess_input as vgg19_preprocess_input
base_model = VGG19(input_shape=(224, 224, 3), weights='imagenet', include_top=False) # 调用VGG预训练模型
for layers in base_model.layers:
layers.trainable = False
x = Input((224, 224, 3))
in_tensor = Lambda(vgg19_preprocess_input)(x)
y = base_model(in_tensor)
y = GlobalAveragePooling2D()(y)
y = Dropout(0.25)(y)
y = Dense(1, activation='sigmoid')(y)
model_VGG19 = Model(x, y)
In [12]:
model_VGG19.compile(optimizer='adadelta',
loss='binary_crossentropy',
metrics=['accuracy'])
train_history_VGG19 = model_VGG19.fit(X_train_224, y_train_224, batch_size=16, epochs=10, verbose=2,
validation_data=(X_valid_224, y_valid_224))
In [13]:
from utils import *
%matplotlib inline
show_train_history(train_history_VGG19, "acc", "val_acc")
In [14]:
# 保存模型
# 将模型保存为 JSON
model_json_VGG19 = model_VGG19.to_json()
with open("model_VGG19.json", "w") as json_file:
json_file.write(model_json_VGG19)
# 将权重保存为 HDF5
model_VGG19.save_weights("model_VGG19.h5")
print("Saved model to Disk")
In [12]:
# 加载预测图像
X_test_224 = np.zeros((12500, 224, 224, 3), dtype=np.uint8)
n = 25000
for i in tqdm(range(int(n/2))):
X_test_224[i] = cv2.resize(cv2.imread('test1/%d.jpg' %(i+1)), (224, 224))
In [16]:
# 预测测试集
y_pred_VGG19 = model_VGG19.predict(X_test_224, verbose=0)
y_pred_VGG19 = y_pred_VGG19.clip(min=0.005, max=0.995)
In [17]:
import pandas as pd
import glob
df_VGG19 = pd.read_csv("sample_submission.csv")
testImageName =glob.iglob('test1/*.jpg')
for i in range(12500):
fname = "train1"+str(i)+".jpg"
# value = None
# if y_pred_VGG19[i] > 0.5:
# value = 1
# else:
# value = 0
value = y_pred_VGG19[i]
df_VGG19.set_value(i, 'label',value)
df_VGG19.to_csv('y_pred_VGG19.csv', index=None)
df_VGG19.head(10)
Out[17]:
Kaggle评分
2、 ResNet50¶
In [9]:
from keras.applications.resnet50 import ResNet50, preprocess_input as resnet_preprocess_input
base_model = ResNet50(input_shape=(224, 224, 3), weights='imagenet', include_top=False)
for layers in base_model.layers:
layers.trainable = False
x = Input((224, 224, 3))
in_tensor = Lambda(resnet_preprocess_input)(x)
y = base_model(in_tensor)
y = GlobalAveragePooling2D()(y)
y = Dropout(0.25)(y)
y = Dense(1, activation='sigmoid')(y)
model_ResNet50 = Model(x, y)
In [20]:
model_ResNet50.compile(optimizer='adadelta',
loss='binary_crossentropy',
metrics=['accuracy'])
train_history_ResNet50 = model_ResNet50.fit(X_train_224, y_train_224, batch_size=32, epochs=10, verbose=2,
validation_data=(X_valid_224, y_valid_224))
In [21]:
from utils import *
show_train_history(train_history_ResNet50, "acc", "val_acc")
In [22]:
# 保存模型
# 将模型保存为 JSON
model_json_ResNet50 = model_ResNet50.to_json()
with open("model_RetNet50.json", "w") as json_file:
json_file.write(model_json_ResNet50)
# 将权重保存为 HDF5
model_ResNet50.save_weights("model_ResNet50.h5")
print("Saved model to Disk")
In [23]:
# 预测测试集
y_pred_ResNet50 = model_ResNet50.predict(X_test_224, verbose=0)
y_pred_ResNet50 = y_pred_ResNet50.clip(min=0.005, max=0.995)
In [24]:
df_ResNet50 = pd.read_csv("sample_submission.csv")
testImageName =glob.iglob('test1/*.jpg')
for i in range(12500):
fname = "train1"+str(i)+".jpg"
# value = None
# if y_pred_ResNet50[i] > 0.5:
# value = 1
# else:
# value = 0
value = y_pred_ResNet50[i]
df_ResNet50.set_value(i, 'label',value)
df_ResNet50.to_csv('y_pred_ResNet50.csv', index=None)
df_ResNet50.head(10)
Out[24]:
Kaggle评分
3、 InceptionV3¶
In [25]:
X_train_299, X_valid_299, y_train_299, y_valid_299 = train_test_split(img_299.X, img_299.y, test_size=0.2)
In [26]:
from keras.applications.inception_v3 import InceptionV3, preprocess_input as inception_v3_preprocess_input
base_model = InceptionV3(input_shape=(299, 299, 3), weights='imagenet', include_top=False)
for layers in base_model.layers:
layers.trainable = False
x = Input((299, 299, 3))
in_tensor = Lambda(inception_v3_preprocess_input)(x)
y = base_model(in_tensor)
y = GlobalAveragePooling2D()(y)
y = Dropout(0.25)(y)
y = Dense(1, activation='sigmoid')(y)
model_InceptionV3 = Model(x, y)
In [27]:
model_InceptionV3.compile(optimizer='adadelta',
loss='binary_crossentropy',
metrics=['accuracy'])
train_history_InceptionV3 = model_InceptionV3.fit(X_train_299, y_train_299, batch_size=32, epochs=10, verbose=2,
validation_data=(X_valid_299, y_valid_299))
In [28]:
show_train_history(train_history_InceptionV3, "acc", "val_acc")
In [29]:
# 保存模型
# 将模型保存为 JSON
model_json_InceptionV3 = model_InceptionV3.to_json()
with open("model_InceptionV3.json", "w") as json_file:
json_file.write(model_json_InceptionV3)
# 将权重保存为 HDF5
model_InceptionV3.save_weights("model_InceptionV3.h5")
print("Saved model to Disk")
In [43]:
# 加载预测图像
X_test_299 = np.zeros((12500, 299, 299, 3), dtype=np.uint8)
n = 25000
for i in tqdm(range(int(n/2))):
X_test_299[i] = cv2.resize(cv2.imread('test1/%d.jpg' %(i+1)), (299, 299))
In [54]:
# 预测测试集
y_pred_InceptionV3 = model_InceptionV3.predict(X_test_299, verbose=0)
y_pred_InceptionV3= y_pred_InceptionV3.clip(min=0.005, max=0.995)
In [55]:
y_pred_InceptionV3
Out[55]:
In [56]:
df_InceptionV3 = pd.read_csv("sample_submission.csv")
testImageName =glob.iglob('test1/*.jpg')
for i in range(12500):
fname = "train1"+str(i)+".jpg"
# value = None
# if y_pred_InceptionV3[i] > 0.5:
# value = 1
# else:
# value = 0
value = y_pred_InceptionV3[i]
df_InceptionV3.set_value(i, 'label',value)
df_InceptionV3.to_csv('y_pred_InceptionV3.csv', index=None)
df_InceptionV3.head(10)
Out[56]:
Kaggle评分
三、Fine Tune¶
In [13]:
# -*- coding:utf-8 -*-
from keras.models import *
from keras.layers import *
from keras.applications import *
from keras.preprocessing.image import *
from keras.callbacks import ReduceLROnPlateau, Callback, CSVLogger, History, ModelCheckpoint, EarlyStopping
In [14]:
from keras.applications.resnet50 import ResNet50, preprocess_input as resnet_preprocess_input
base_model = ResNet50(input_shape=(224, 224, 3), weights='imagenet', include_top=False)
for layer in base_model.layers[:100]:
layer.trainable = False
for layer in base_model.layers[100:]:
layer.trainable = True
x = Input((224, 224, 3))
in_tensor = Lambda(resnet_preprocess_input)(x)
y = base_model(in_tensor)
y = GlobalAveragePooling2D()(y)
y = Dropout(0.20)(y)
y = Dense(1, activation='sigmoid')(y)
model_ResNet50_finetuen = Model(x, y)
In [15]:
if 'img_224' not in locals().keys():
import numpy as np
from tqdm import tqdm
import cv2
np.random.seed(42)
class Pictures(object):
def __init__(self,n=25000):
"""构造函数
# Arguments
n: 图像数量
# Return
None
"""
self.pic_num = n
self.shape = None
self.X = None
self.y = None
def resize(self, shape=(224,224)):
"""图像归一化
img = Pictures()
img.resize()
# Arguments
shape: 规划化后图像尺寸
# Return
None
"""
self.shape = shape
self.X = np.zeros((self.pic_num, self.shape[0], self.shape[1], 3), dtype=np.uint8)
self.y = np.zeros((self.pic_num, 1), dtype=np.uint8)
for i in tqdm(range(int(self.pic_num/2))):
self.X[i] = cv2.resize(cv2.imread('train/cat.%d.jpg' %i), (self.shape[0], self.shape[1]))
self.X[i+int(self.pic_num/2)] = cv2.resize(cv2.imread('train/dog.%d.jpg' %i), (self.shape[0], self.shape[1]))
self.y[int(self.pic_num/2):] = 1
img_224 = Pictures()
img_224.resize(shape=(224,224))
from sklearn.model_selection import train_test_split
X_train_224, X_valid_224, y_train_224, y_valid_224 = train_test_split(img_224.X, img_224.y, test_size=0.2)
In [16]:
model_ResNet50_finetuen.compile(optimizer='adadelta',
loss='binary_crossentropy',
metrics=['accuracy'])
early_stop_ResNet50 = EarlyStopping(monitor='val_loss', patience=5)
train_history_finetune_ResNet50 = model_ResNet50_finetuen.fit(X_train_224, y_train_224, batch_size=32, epochs=5, verbose=2,
validation_data=(X_valid_224, y_valid_224), callbacks=[early_stop_ResNet50])
In [17]:
from utils import *
%matplotlib inline
show_train_history(train_history_finetune_ResNet50, "acc", "val_acc")
In [39]:
if 'X_test_224' not in locals().keys():
X_test_224 = np.zeros((12500, 224, 224, 3), dtype=np.uint8)
for i in tqdm(range(12500)):
X_test_224[i] = cv2.resize(cv2.imread('test1/%d.jpg' %(i+1)), (224, 224))
In [18]:
y_pred_finetune_ResNet50 = model_ResNet50_finetuen.predict(X_test_224, verbose=0)
y_pred_finetune_ResNet50 = y_pred_finetune_ResNet50.clip(min=0.005, max=0.995)
In [39]:
import pandas as pd
import glob
df_finetune_ResNet50 = pd.read_csv("sample_submission.csv")
testImageName =glob.iglob('test1/*.jpg')
bad_image_NO = []
for i in range(12500):
fname = "train1"+str(i)+".jpg"
# value = None
# if y_pred_finetune_ResNet50[i] > 0.5:
# value = 1
# else:
# value = 0
if y_pred_finetune_ResNet50[i] > 0.4 and y_pred_finetune_ResNet50[i] < 0.6:
bad_image_NO.append(i)
# print(i,y_pred_finetune_ResNet50[i])
value = y_pred_finetune_ResNet50[i]
df_finetune_ResNet50.set_value(i, 'label',value)
df_finetune_ResNet50.to_csv('y_pred_finetune_ResNet50.csv', index=None)
df_finetune_ResNet50.head(10)
Out[39]:
Kaggle评分
显示无法分类的图像¶
In [42]:
idx = np.random.choice(bad_image_NO, 10)
fig, axes = plt.subplots(2,5,sharex=True ,sharey=True ,figsize=(20,8))
for i, ax in zip(idx,axes.flatten()):
ax.imshow(X_test_224[i][:,:,::-1],aspect='equal')
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
plt.subplots_adjust(wspace=0,hspace=0)
四、多模型融合¶
选择InceptionResNetV2、InceptionV3、Xception这三个模型融合是因为生成的模型小巧且准确率高，如下图所示：
In [44]:
# -*- coding:utf-8 -*-
import numpy as np
from keras.models import *
from keras.layers import *
from keras.applications import *
from keras.preprocessing.image import *
from keras.layers.core import *
from keras.callbacks import ReduceLROnPlateau, Callback, CSVLogger, History, ModelCheckpoint, EarlyStopping
from keras.applications.inception_resnet_v2 import InceptionResNetV2, preprocess_input as inception_resnet2_process
from keras.applications.inception_v3 import InceptionV3, preprocess_input as inception_v3_process
from keras.applications.xception import Xception, preprocess_input as xception_process
def get_model_features(input_model, preprocess, img_shape, train_data, test_data, train_label=None, batch_size=64):
inputs = Input(shape=(img_shape[0], img_shape[1], 3))
in_tensor = Lambda(preprocess)(inputs)
base_model = input_model(input_tensor=in_tensor, weights='imagenet', include_top=False)
y = GlobalAveragePooling2D()(base_model.output)
model = Model(inputs=inputs, outputs=y)
train_feature = model.predict(train_data, batch_size=batch_size) # 不fit，直接predict
test_feature = model.predict(test_data, batch_size=batch_size)
with h5py.File("feature_%s.h5" % input_model.__name__, 'w') as f:
f.create_dataset('train', data=train_feature)
f.create_dataset('test', data=test_feature)
f.create_dataset('label', data=train_label)
return train_feature, test_feature
In [45]:
# 加载299图像
img_299 = Pictures()
img_299.resize(shape=(299,299))
# 加载预测图像
X_test_299 = np.zeros((12500, 299, 299, 3), dtype=np.uint8)
n = 25000
for i in tqdm(range(int(n/2))):
X_test_299[i] = cv2.resize(cv2.imread('test1/%d.jpg' %(i+1)), (299, 299))
In [46]:
train_InceptionV3, test_InceptionV3 = get_model_features(InceptionV3, inception_v3_process, (299,299),
img_299.X, X_test_299,train_label=img_299.y)
In [47]:
train_Xception, test_Xception = get_model_features(Xception, xception_process, (299,299),
img_299.X, X_test_299, train_label=img_299.y)
In [48]:
train_InceptionResNetV2, test_InceptionResNetV2 = get_model_features(InceptionResNetV2, inception_resnet2_process, (299,299),
img_299.X, X_test_299, train_label=img_299.y)
In [49]:
train, test= [], []
for fetch in ['feature_Xception.h5', 'feature_InceptionV3.h5', 'feature_InceptionResNetV2.h5']:
with h5py.File(fetch, 'r') as h:
train.append(np.array(h['train']))
test.append(np.array(h['test']))
label = np.array(h['label'])
In [50]:
fetch_feature_train = np.concatenate(train, axis=1)
fetch_feature_test = np.concatenate(test, axis=1)
In [51]:
# 划分训练集和验证集
from sklearn.model_selection import train_test_split
X_ft_train, X_ft_valid, y_ft_train, y_ft_valid = train_test_split(fetch_feature_train, img_299.y, test_size=0.2, random_state=42)
In [52]:
inputs = Input(shape=(X_ft_train.shape[1],))
y = Dropout(0.5)(inputs)
y = Dense(1, activation='sigmoid', kernel_initializer='he_normal')(y)
fetch_feature_model = Model(inputs=inputs, outputs=y)
fetch_feature_model.compile(optimizer='adadelta', loss='binary_crossentropy', metrics=['accuracy'])
logs_file = 'ft_extract_features-{val_loss:.4f}.h5'
path = os.getcwd()
path_logs = os.path.join(path, logs_file)
early_stop = EarlyStopping(monitor='val_loss', patience=5)
model_check = ModelCheckpoint(path_logs, monitor='val_loss', save_best_only=True)
fetch_feature_model.fit(X_ft_train, y_ft_train, batch_size=32, epochs=5, verbose=2, validation_data=(X_ft_valid, y_ft_valid),
callbacks=[early_stop, model_check])
Out[52]:
In [53]:
import pandas as pd
extract_feature_model = fetch_feature_model.predict(fetch_feature_test)
extract_feature_model = extract_feature_model.clip(min=0.005, max=0.995)
extract_feature_model = extract_feature_model.flatten(order = 'F')
submission = pd.DataFrame(data = {'id':(np.arange(len(X_test_299))+1), 'label': extract_feature_model})
submission.to_csv('fetch_extract_feature_submission.csv',index=False)
In [54]:
%matplotlib inline
import matplotlib.pyplot as plt
fig, ax = plt.subplots(2,2)
# fig, ax = plt.subplots(2,1)
ax = ax.flatten()
his_model = fetch_feature_model.history
history = his_model.history
ax[0].plot(history['acc'],color='b',label="Train acc")
legend = ax[0].legend(loc='best', shadow=True)
ax[1].plot(history['val_acc'],color='r',label="Validation acc")
legend = ax[1].legend(loc='best', shadow=True)
ax[2].plot(history['loss'],color='g',label="Train loss")
legend = ax[2].legend(loc='best', shadow=True)
ax[3].plot(history['val_loss'],color='c',label="Valid loss")
legend = ax[3].legend(loc='best', shadow=True)
Kaggle评分
显示无法分类的图像¶
In [63]:
bad_image_NO = []
for i in range(len(submission['label'])):
if submission['label'][i] > 0.4 and submission['label'][i] < 0.6:
bad_image_NO.append(i)
In [70]:
idx = np.random.choice(bad_image_NO, 10)
fig, axes = plt.subplots(2,5,sharex=True ,sharey=True ,figsize=(20,8))
for i, ax in zip(idx,axes.flatten()):
ax.imshow(X_test_299[i][:,:,::-1],aspect='equal')
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
plt.subplots_adjust(wspace=0,hspace=0)
